# The Black-Litterman Approach

:
Original Model and Extensions1 Attilio Meucci2
attilio_meucci@symmys.com

This version: March 2010 Last version available at www.ssrn.com

Abstract We walk the reader through the Black-Litterman approach, providing all the proofs. We show how minor modiﬁcations of the original model greatly improve its range of applications. We discuss full generalizations of this and related models. Code is available at MATLAB Central File Exchange. JEL Classiﬁcation: C1, G11

1 A shorter version of this article appears as Meucci, A., 2010, The Black-Litterman Approach: Original Model and Extensions, The Encyclopedia of Quantitative Finance, Wiley 2 The author gratefully acknowledges the very helpful feedback from Bob Litterman and Jay Walters

1

see also Satchell and Scowcroft (2000). To set π. BL acknowledge and address the issue estimation risk: since µ cannot be known with certainty. it is modeled as a random variable whose dispersion represents the possible estimation error. the expected returns on those portfolios. In Section 3 we rephrase BL in terms of views on the market. Idzorek (2004) and Walters (2008). the views are normal.1 Introduction At a time when portfolio optimization used to take as inputs only the expectations and the covariances of a set of assets computed from a given reference econometric model. the conﬁdence in the view portfolios and the uncertainty on the reference model. Σ) . views on external factors that inﬂuence the p&l indirectly through correlation. i. To specify µ. views on generalized risk factors rather than returns. In Section 2 we review the original BL methodology: the reference model is normal. The market model We consider a market of N securities or asset classes.e. In particular. and the distribution that blends these two inputs is obtained analytically using Bayes’ formula. Σ) . a portfolio manager could process those inputs. τ ≡ 0 in (2). and obtain an optimal allocation that reﬂected the views in a consistent way without corner solutions. 2 (3) . the pathbreaking technique by Black and Litterman (1990) (BL in the sequel) provided a framework in which more satisfactory results could be obtained from a larger set of inputs: view portfolios. 2 The original model Here we follow He and Litterman (2002). stress-test of correlations and volatilities. (2) where π represents the best guess for µ and τ Σ the uncertainty on this guess. blend them into the reference return distribution. centered around the CAPM equilibrium. (1) The covariance Σ is estimated by exponential smoothing of the past return realizations. Assuming there is no estimation error. non-normal markets. Using BL. τ Σ) . markets of complex derivatives. instead of the market parameters: this market-based version is more parsimonious and it allows for the inclusion of scenario analysis as a special case. In Section 4 we review the literature related to BL and its extensions: ranking views. the reference model (1) becomes X ∼ N (π. whose returns are normally distributed: X ∼ N (µ. multiple users. BL invoke an equilibrium argument. BL state that µ is normally distributed µ ∼ N (π.

portfolio optimization is extremely sensitive to the input parameters. where BL set exogenously λ ≈ 1. but extra parameters need to be calibrated. we can compare the speciﬁcation (2) with the uncertainty on the sample estimator (6) and set 1 τ≈ . (8) T Satchell and Scowcroft (2000) propose an ingenious model where τ is stochastic. π can be set in terms of w. 1) is called for in most applications. w (4) Setting to zero the derivative with respect to w of the term in curly brackets we can solve explicitly this problem. One way to cope with this issue is to use as in Stein (1955) more eﬃcient "shrinkage" estimators: b µ(s) ≡ (1 − s) µ + sπ 0 . Britten-Jones (1999) and Meucci (2005) for a review. (6) T t=1 T e π ≡ 2λΣw. a tailor-made calibration that spans the interval (0. see also the discussion in Walters (2008).2. In practice. Therefore. Indeed. Giacometti. Green and Holliﬁeld (1992). and Fabozzi (2007) generalize this argument to stable-distributed markets. Chopra and Ziemba (1993).Assume that. In BL the risk drivers X represent the returns of a broad market that are independently and identically normally distributed across time as in (3). (7) where π 0 is a generic shrinkage target and 0 ≤ s ≤ 1 is the amount of shrinkage. see Section 4. The sample mean is a very ineﬃcient estimator: therefore. To calibrate the overall uncertainty level τ in the reference model. 3 . Rachev. The standard estimator of the expectations in this context is the sample mean µ ¶ T 1X Σ b µ≡ Xt ∼ N π. Modiﬁcations of BL with diﬀerent targets were considered early on by the authors. all investors maximize a mean-variance trade-oﬀ and that the optimization is unconstrained: wλ ≡ argmax {w0 π − λw0 Σw} . . (5) where T is the length of the available time series. in particular the expected values. Best and Grauer (1991). the "optimal" allocations based on (6) vary wildly when diﬀerent time series are fed into the allocation process. Notice that historical information does not play a direct role in the determination of π: this is an instance of the shrinkage approach to estimation risk. Bertocchi. see Jobson and Korkie (1980). obtaining the relationship between the e equilibrium portfolio w which stems from an average risk-aversion level λ and the reference expected returns: e Therefore. consistently with this normal market. the BL prior can be seen as an extreme case of shrinkage toward the theoretical expectations (5) implied by equilibrium.

. In order to associate uncertainty with the views. K. For instance. Spain. US and Germany. 8%)0 and obtain from (5) the annualized expected 0 returns π ≈ (6%. . 8%. c (12) where the meta-parameters v and Ω quantify views and uncertainty thereof respectively. . Switzerland. it is convenient to set the entries of v in terms of the volatility induced by the market: q¡ ¢ vk ≡ (Pπ)k + η k (11) PΣP0 k.4 in (2).To determine the prior expectation π we start from the market-weighted portfoe lio w ≈ (4%. −α. 29%. "bearish". 71%. 17%. BL consider views on expectations. 8%. 4%. Furthermore. whose generic k-th row determines the relative weight of each expected return in the respective view. 10%) . "bullish" and "very bullish" views respectively. k = 1. If the user has only qualitative views. this corresponds to statements on the parameter µ. 4 . 9%. 7%. where the scatter structure of uncertainty is inherited from the market volatilities and correlations and c represents an overall level of conﬁdence in the views. we set τ ≈ 0. Finally. Also. 5%. . BL use a normal model: Pµ ∼ N (v. 31%) and the correlation matrix ⎞ ⎛ 1 54% 62% 25% 41% 59% ⎜ · 1 69% 29% 36% 83% ⎟ ⎟ ⎜ ⎜ · · 1 15% 46% 65% ⎟ ⎟ ⎜ (9) C≈⎜ · · 1 47% 39% ⎟ ⎟ ⎜ · ⎝ · · · · 1 38% ⎠ · · · · · 1 where ηk ∈ {−β. Typical choices for these parameters are α ≡ 1 and β ≡ 2. The covariance matrix of daily returns on the above classes Σ is estimated as follows in terms of the (annualized) volatilities σ ≈ (21%. In the normal market (1). The views A view is a statement on the market that can potentially clash with the reference market model (1). it is convenient to set as in Meucci (2005) Ω≡ 1 PΣP0 .k . in which case the view is X3 − X2 ≥ 0. Canada. we consider the oversimpliﬁed case of an international stock fund that invests in the following six stock market national indexes: Italy. +β} deﬁnes "very bearish". Ω) . Another view could be that the fourth asset class experiences twice to three times the volatility predicted by the model: 4Σ44 ≤ Var {X4 } ≤ 9Σ44 . 24%. BL focus on linear views: K views are represented by a K × N "pick" matrix P. 24%. the portfolio manager might say that the third asset class will outperform the second. (10) To illustrate. 25%. +α.

(14) BL where ³ ´−1 ³ ´ −1 −1 µBL ≡ (τ Σ) + P0 Ω−1 P (τ Σ) π + P0 Ω−1 v ³ ´−1 Σµ ≡ (τ Σ)−1 + P0 Ω−1 P . (17) where µBL is deﬁned in (15) and ΣBL follows from (16) by assuming that µ and Z are independent: ΣBL ≡ Σ + Σµ . possibly under a set of linear constraints. (19) (20) The normal posterior distribution (17) with (19) and (20) represents the modiﬁcation of the reference model (3) that incorporates the views (10). we are interested in the distribution of the risk factors X. c ≡ 1 in (12). The ensuing eﬃcient frontier represents a gentle twist to equilibrium that reﬂects the views without extreme corner solutions. such as boundaries on securities/asset classes.To continue with our example.2. Ω ∼ N (µBL . ΣBL ) . BL (15) 0 and (16) However.e. where d Z ∼ N (0. (18) BL As we prove in Appendix 5. Therefore the pick matrix reads µ ¶ 0 1 0 0 0 0 P≡ . (13) 0 0 0 0 1 −1 and the annualized views vector becomes v ≡ (12%. The posterior In Appendix 5. or a budget constraint. This quadratic programming problem can be easily solved numerically. Σ). 5 . the manager might assess two views: the Spanish index will rise by 12% on an annualized basis. Therefore the posterior market model is X|v.1 we show how to obtain the distribution of µ given the views using Bayes’ formula: µ|v. Ω = µ|v. The allocation With the posterior distribution it is now possible to set and solve a meanvariance optimization. Σµ ) . Ω ∼ N (µBL . and the spread USGermany will experience a negative annualized change of 10%. −10%) . To d compute this distribution we rewrite the reference model (1) as X = µ+Z. We set the uncertainty in the views of the same order of magnitude as the market. Ω + Z or X|v. i. computationally more stable representation of the posterior parameters (15) and (18) reads: µBL ΣBL ¢−1 ¡ = π + τ ΣP0 τ PΣP0 + Ω (v − Pπ) ¢−1 ¡ 0 0 2 = (1 + τ ) Σ − τ ΣP τ PΣP + Ω PΣ. an equivalent.

3 The market formulation The BL posterior distribution (17) presents two puzzles. Σ) .0 1 8 volatility POSTERIOR MODEL B L v i e w s 1 composition 0 . 0 2 US volatility Figure 1: BL: eﬃcient frontier twisted according to the views In our example we assume the standard long-only and budget constraints. 3 0 . 0 1 8 0 . i.0 1 9 0 . 9 0 .REFERENCE MODEL r e f e r e n c e 1 0 . and the exposure to the US market decreases. Spain Italy 0 . 4 0 . In Figure 1 we plot the eﬃcient frontier from the reference model (3) and from the posterior model (17). 0 1 5 0 . 5 0 . 2 0 .0 1 7 0 . In BL. the covariance of the uninformative posterior model appears distorted. 3 0 . (22) In other words. one would expect the posterior to equal the reference model (3).e. 5 0 .0 1 6 0 . 9 Germany composition 0 . 0 1 4 0 . the model does not depart from the reference prior. the reference model (21) was derived by assuming no estimation risk.0 1 5 0 . Spain Italy 0 . 7 0 . unless τ ≡ 0. 0 1 4 0 .e. 1 0 Germany Canada Switz. Ω → ∞ in (10). the exposure to the Spanish market increases for lower risk values. 6 0 . w ≥ 0 and w0 1 ≡ 1. i. 0 1 3 US 0 0 . τ ≡ 0.e. when the views are uninformative. 6 . 0 1 6 0 . 7 0 . Indeed. 6 0 . i. (1 + τ ) Σ) . (21) This condition is important to ensure that. when the user does not have strong views. in this limit the posterior becomes X ∼ N (π. 0 1 7 0 . 0 1 3 0 . On one extreme. 8 0 . 2 0 . 4 0 . which we report here: X ∼ N (π. 0 1 1 0 . This result is actually fully consistent. 8 0 . Consistently with the views. 1 Canada Switz. the exposure to Germany increases across all levels of risk aversion. 0 1 2 0 .

which represents the estimation error on µ. Σ|v) . The user has views on linear functions of the market V ≡ PX. . To ﬁx this problem we can proceed as in Meucci (2005) and rephrase BL in terms of views directly on the market X. To summarize. (26) BL BL where µΩ=0 BL ΣΩ=0 BL ¡ ¢−1 = π + ΣP0 PΣP0 (v − Pπ) ¡ ¢−1 = (1 + τ ) Σ − τ ΣP0 PΣP0 PΣ. although the null.e. . Therefore we set µ ≡ π as in (3) to obtain the reference model without estimation risk (21). one would expect the posterior to become the reference model (21) conditioned on the speciﬁc views. the expectation of the full-conﬁdence posterior (27) equals the conditional expectation of scenario analysis (24). where µ|v Σ|v ¡ ¢−1 ≡ π + ΣP0 PΣP0 (v − Pπ) ¡ ¢ 0 0 −1 ≡ Σ − ΣP PΣP PΣ. but the full-conﬁdence posterior covariance (28) never yields the conditional covariance of scenario analysis (25): the volatility-correlation portion Σ of the covariance in (18) prevents the BL full-conﬁdence posterior covariance from ever becoming degenerate. not on the market X: therefore. they might clash with the practitioner’s intuition. in addition to the pure volatility-correlation structure of the market Σ. . in Mina and Xiao (2001). i. when the conﬁdence in the views v is full. where P is a pick matrix as in (10). (24) (25) (23) The conditional distribution (23) is the core of scenario analysis: the user selects 0 a set of deterministic scenarios v ≡ (v1 . the views (10) are expressed on the parameter µ. the conﬁdence in the views is only supposed to aﬀect the estimation risk part of the covariance in (22). On the other extreme. Again. (24)-(25) generalize the classical regression-like result utilized e. Indeed. the full-conﬁdence posterior becomes ¡ ¢ X ∼ N µΩ=0 . the reference model would coincide exactly with the uninformative model (22). ΣΩ=0 . which. Ω → 0. as proved in Meucci (2005). we do not consider µ as a random variable. However. However. (27) (28) Therefore. the covariance of (22) is not as easy to interpret because it contains. the additional term τ Σ. . In BL. is also normal: X|v ∼ N (µ|v.g.Without that assumption. vK ) for the combinations of factor realizations that are assumed to take place and analyzes their eﬀect on the reference model. this apparent paradox can be explained: in BL. not its volatility-correlation component Σ. instead of the parameter µ. The distribution of 7 . We start from the reference model (1).and full-conﬁdence limits of the BL posterior distribution are fully consistent.

Ω → 0. As a result. i. However.e. At the other extreme. Once the model is set up.3. it is immediate to check that if the conﬁdence is null. i. Therefore. see . (31) (32) (30) These market-posterior formulas are very similar to their counterparts (19)-(20) in the original BL model. if the conﬁdence in the views is full. in BL there is no historical updating of the estimate of µ 8 . the user quantiﬁes his views. Σm ) .the views in general clashes with the distribution implied by the reference model (21): as a random variable. BL adds uncertainty on the views by means of Bayesian formulas3 . in the market based-speciﬁcation there is no need to add the original covariance to the posterior as in (18). in the ﬁrst place the parameter τ in (2) never appears here. and Klein (1979) and Goel and Zellner (1986): in particular. which is also normal X|v. Σm ) BL BL mkt-posterior X ∼ N (µ|v. Qian and Gorman (2001) use a conditional/marginal factorization to input views on 3 However. this is not the standard Bayesian framework of decision theory. BL BL where µm BL Σm BL ¢−1 ¡ ≡ π + ΣP0 PΣP0 + Ω (v − Pπ) ¢−1 ¡ 0 0 ≡ Σ − ΣP PΣP + Ω PΣ. as it turns out. Brown. see Bawa. % & X ∼ N (π. choosing a speciﬁc value v for the random variable V. Σ|v) conditional (full conﬁdence: Ω → 0) 4 Related literature Scenario analysis (23) allows the practitioner to explore the implications on a given portfolio of a set of subjective views on possible market realizations. (29) where Ω represents the uncertainty on the views as in (10).e. then the posterior (30) equals the conditional distribution (23). Ω → ∞. Second. Σ) prior (no conﬁdence: Ω → ∞) X ∼ N (µm . Indeed. then the posterior (30) equals the prior (21). the market speciﬁcation is consistent with both the reference model and with scenario analysis. the views V as a random variable are a perturbation of the outcome implied by the reference model and as such they are modeled as a conditional distribution V|x ∼ N (Px. the realization of V according to the views could turn out larger or smaller than the realization of PX according to the reference model. applying Bayes’ rule as in the original BL we obtain the posterior distribution of the market given the views. Ω ∼ N (µm . As we show in Appendix 5. Ω) .

but with diﬀerent probabilities: therefore. 4 Indeed. "lax" views on expectations. Instead. active management was among the ﬁrst applications of BL by their authors. although it does not cover all the above desired applications. the posterior formulas (19)-(20) and their modiﬁed versions (31)-(32). it is more natural to input views directly on the market. We summarize in the table below the capabilities of the original BL and its market formulation in Section 3. thereby handling views and stress-testing in derivatives markets. Meucci (2009). The COP approach in Meucci (2006) proceeds in this direction. whereas opinion pooling accounts for diﬀerent conﬁdence levels and multiple users. tail behaviors.. not necessarily equilibrium4 . Pezier (2007). The entropy pooling approach in Meucci (2008) solves the above issues: a posterior consistent with the most general views is obtained naturally by entropy minimization. ranges. Pezier (2007) processes full and partial views on expectations and covariances by least discrimination. because the practitioner wishes to express views on the expectations of the market and. under the normal assumption.volatilities and correlations in addition to expectations. where portfolios were constrained to be orthogonal to the market and the prior expectation was assumed null. instead of securities returns. Further generalizations of BL should handle non-normal reference risk models. Furthermore. as well as the formulas in the above literature can be applied to any normal distribution. non-linear views. the COP in Meucci (2006) and the entropy-pooling approach in Meucci (2008). the Bayesian approach in BL acts on the parameter µ of a normal distribution: this is correct. etc. and simultaneous inputs from multiple users with diﬀerent hierarchical conﬁdence levels. as no costly repricing is ever necessary. 9 . the COP relies on ad-hoc manipulations. Accordingly. However. even the most complex derivatives can be handled. views on generalized features such as medians. in non-normal markets with views on features other than expectations one should develop a diﬀerent approach for each diﬀerent possible market parametrization and for each possible feature on which the practitioner wishes to express views. equality and inequality (ranking) views. these are represented by µ. instead of the combinations of parameters that correspond to those features. However. The equilibrium-based prior expectations (5) in the original BL and in the above literature appears to restrict the potential applications of these approaches to the tactical management of a global diversiﬁed fund. Qian and Gorman (2001). Almgren and Chriss (2006). the posterior is represented in terms of the same Monte Carlo scenarios as the reference model. Almgren and Chriss (2006) provide a framework to express ranking. More importantly. Meucci (2009) applies the above approaches to fully generic risk factors that map non-linearly into the ﬁnal p&l. The Bayesian formalism of BL cannot achieve this: aside from the insurmountable computational problems when leaving the normal assumption.

Estimation Risk and Optimal Porfolio Choice (North Holland). and R. 655—671. When will mean-variance eﬃcient portfolios be well diversiﬁed?. Bertocchi. Journal of Finance 47. and R. J. Chriss. Chopra. and R.M.S. R.. 1990. The eﬀects of errors in means. Optimal portfolios from ordering information. Review of Financial Studies 4. Journal of Risk 9. Bayesian Inference and Decision Techniques: Essays in Honor of Bruno De Finetti . M. V. M. Journal of Finance 54. Best. Ziemba. Britten-Jones. and A. vol. Grauer. S. Journal of Portfolio Management pp. and W.normal market & linear views scenario analysis correlation stress-test trading desk: non-linear pricing external factors: macro. Stable distributions in the Black-Litterman approach to asset allocation. M. and B. 2002. Litterman. T. Fabozzi. ssrn. Zellner. etc. On the sensitivity of mean-variance-eﬃcient portfolios to changes in asset means: Some analytical and computational results. 10 . Giacometti. 1999. and covariances on optimal portfolio choice. 1993. 6 of Studies in Bayesian Econometrics and Statistics (Elsevier Science). T.. C. 1785—1809. and F. 315—342. J. He. W. V. The sampling error in estimates of mean-variance eﬃcient portfolio weights. Bawa. 6—11.. 423—433. S. Goldman Sachs Fixed Income Research.com. Litterman. R. 1991. Goel. and N. 1979. G.. Rachev. Black... Brown. 2006. The intuition behind Black-Litterman model portfolios. 2007. Green. Holliﬁeld. I. F. Asset allocation: combining investor views with market equilibrium.. 1992. Quantitative Finance 7. P.. 1—47. 1986.. R.. partial speciﬁcations non-normal market multiple users non-linear views trading desk: costly pricing lax constraints: ranking BL X X · · · · · · · · · AC · · · · · · · · · · X QG X X X · · · · · · · · P X X X · · X · · · · · M X X X X X X · · · · · COP X X · X X · X X · · · EP X X X X X X X X X X X References Almgren. and R. Klein. variances.

. 114—119. Estimation for Markowitz eﬃcient portfolios. Pezier. Beyond Black-Litterman in practice: A ﬁve-step recipe to input views on non-normal markets. RiskMetrics publications. Satchell. J. 138—150.. 1980.. Journal of Asset Management 10. Fully ﬂexible views: Theory and practice. D.. and B. . Enhancing the Black-Litterman and related approaches: Views and stress-test on risk factors. and J. Mina. 89—96 Available at www. Journal of Asset Management 1. 2008.. 2001. Journal of the American Statistical Association 75.com. Return to RiskMetrics: The evolution of a standard. and A. 11 . 2001. Risk and Asset Allocation (Springer). 2004.com.. A demystiﬁcation of the Black-Litterman model: Managing quantitative and traditional construction. Financial Analyst Journal 57.Idzorek. Global portfolio optimization revisited: A least discrimination alternantive to Black-Litterman. The Black-Litterman model: A detailed exploration. 2008. blacklitterman. Qian. 2006. Stein. and S. J.Y. . Jobson. 44—51.org. C. A step-by-step guide to the Black-Litterman model. Risk 19. A.. Proceedings of the 3rd Berkeley Symposium on Probability and Statistics. Zephyr Associates Publications.M. 97—102 Available at www.ssrn. S.. Gorman. Walters. J. 2009. E. Risk 21. 1955. T. Meucci. 2000. J. 544—554. . Xiao. Conditional distribution in portfolio theory. 2005.ssrn.. 2007. Scowcroft. ICMA Centre Discussion Papers in Finance. Korkie. Inadmissibility of the usual estimator for the mean of a multivariate normal distribution.

=R fV (v) fV|µ (v) fµ (µ) dµ (37) Our strategy will be to prove that the numerator can be written as a conditional pdf fV|µ times a factor: the former will be the required pdf. The conditional pdf of this variable is thus fV|µ (v) ≡ |Ω|− 2 1 (35) (2π) K 2 e− 2 (v−Pµ) Ω 1 0 −1 (v−Pµ) . Σµ ) .V (µ. which refers to the BLm model. v) . Ω) . reads: fµ. 5. (33) (34) where Z ∼ N (0.e. BL 12 (40) . d e− 2 (µ−π) (τ Σ) 1 0 −1 (µ−π) . once the following substitutions are made: BL µ π Στ ←→ ←→ ←→ BLm X µ Σ (39) Therefore from (57). we can model v as a random variable V whose distribution. Ω).5 Technical Appendix In this appendix we discuss some technical results that can be skipped at ﬁrst reading. i.1 Derivation of BL |(τ Σ)|− 2 (2π) N 2 1 Here we adapt from Satchell and Scowcroft (2000). we do not care about the latter. This formula is exactly like (52) below. v) = fV|µ (v) fµ (µ) ∝ |(τ Σ)| −1 2 (38) e − 1 [(µ−π)0 (τ Σ)−1 (µ−π)+(v−Pµ)0 Ω−1 (v−Pµ)] 2 |Ω| −1 2 . Ω ∼ N (µBL . (54) and (58) we obtain µ|v. (36) To determine the posterior of µ given V we apply Bayes rule fµ|v (µ) = fV|µ (v) fµ (µ) fµ. The numerator in (37).V (µ. Therefore. The pdf of (2) is fµ (µ) ≡ As for (10) we write it as v = Pµ + Z. the joint pdf of µ and V. conditioned on the realization of µ reads: V|µ ∼ N (Pµ.

using (43) we write (18) as ΣBL ³ ´−1 −1 ≡ Σ + (τ Σ) + P0 Ω−1 P (47) −1 −1 (v − Pπ) (46) = (1 + τ ) Σ − (τ Σ) P0 (P (τ Σ) P0 + Ω) P (τ Σ) 5.where µBL Σµ BL ´−1 ³ ´ −1 + P0 Ω−1 P (τ Σ) π + P0 Ω−1 v ³ ´−1 −1 ≡ (τ Σ) + P0 Ω−1 P . Σ) . (45) which can be easily checked left-multiplying both sides by (P (τ Σ) P0 + Ω). we can write (44) as µBL = π + (τ Σ) P0 (P (τ Σ) P0 + Ω) Similarly.2 Reshuﬄing BL expressions Using the following matrix identity (A and D invertible. (48) Therefore fX (x) ≡ |Σ|− 2 1 (2π) N 2 e− 2 (x−π) Σ 1 0 −1 (x−π) . conformable. Noticing that Ω−1 − (P (τ Σ) P0 + Ω) P (τ Σ) P0 Ω−1 = (P (τ Σ) P0 + Ω) −1 . ³ ´³ ´ −1 = (τ Σ) − (τ Σ) P0 (P (τ Σ) P0 + Ω) P (τ Σ) (τ Σ)−1 π + P0 Ω−1 v ³ ´ −1 = π + (τ Σ) P0 Ω−1 − (P (τ Σ) P0 + Ω) P (τ Σ) P0 Ω−1 v − (τ Σ) P0 (P (τ Σ) P0 + Ω) −1 −1 Pπ. B and C generic.3 Derivation of BLm This proof is adapted from Meucci (2005). ≡ (τ Σ) −1 ³ (41) (42) 5. We report here the reference model (21) for the market X ∼ N (π. potentially of diﬀerent size) ¡ ¢−1 ¡ ¢−1 A − BD−1 C = A−1 − A−1 B CA−1 B − D CA−1 (43) we can write (15) as µBL ≡ ³ ´−1 ³ ´ −1 −1 (τ Σ) + P0 Ω−1 P (44) (τ Σ) π + P0 Ω−1 v . (49) 13 .

In Appendix 5.2 for the regular BL. (50) (2π) To determine the posterior of the market given the views we apply Bayes’ rule fX|v.V (x. v) . BL BL where µm (v) is deﬁned in (54) and BL ¡ ¢−1 Σm ≡ Σ−1 + P0 Ω−1 P . i. The numerator in (51). BL fX. as we did in Appendix 5. (54) where e and v does not depend on either x or v.Also. v) = fV|x (v) fX (x) ∝ |Σ| −1 2 (52) [(x−π) Σ 0 −1 |Ω| −1 2 e −1 2 (x−π)+(v−Px) Ω 0 −1 (v−Px)] . BL (58) (57) Now we proceed to reshuﬄe the expressions (54) and (58) in a more friendly and computationally eﬃcient way. =R fV (v) fV|x (v) fX (x) dx (51) Our strategy will be to prove that the numerator can be written as a conditional pdf fX|v . we do not care about the latter. the joint pdf of V and X. Therefore (55) where the conditional pdf of the market given the views is ¯ ¯1 fX|v (x|v) ∝ ¯Σ−1 + P0 Ω−1 P¯ 2 m 0 −1 0 −1 m 1 e− 2 (x−µBL ) (Σ +P Ω P)(x−µBL ) (56) Since (56) is a normal probability density function. times a factor that only depends on v: the former will be the required pdf. v) ∝ ¯Σ−1 + P0 Ω−1 P¯ 2 e− 2 (x−µBL ) (Σ +P Ω P)(x−µBL ) (53) −1 ¯ ¯ 1 ¯Ω + PΣP0 ¯− 2 e− 1 (v−v)0 (Ω+PΣP0 ) (v−v) .Ω (x) = fV|x (v) fX (x) fX. 2 ¡ ¢−1 ¡ −1 ¢ µm ≡ Σ−1 + P0 Ω−1 P Σ π + P0 Ω−1 v . Σm ) . v) ∝ fX|v (x|v) g (v) .V (x.e.4 below we show that ¯ ¯1 m 0 −1 0 −1 m 1 fX.V (x. 14 . from (29) we obtain fV|x (v) ≡ |Ω| −1 2 K 2 e− 2 (v−Px) Ω 1 0 −1 (v−Px) . we obtain for the posterior market conditioned on the views X|v ∼ N (µm .V (x. reads: fX.

(63) 5. BL (65) or equivalently (66) Using (66) we easily re-write (64) as follows: ¢ ¢ ¡ ¡ [· · · ] = x0 Σ−1 + P0 Ω−1 P x − 2x0 Σ−1 + P0 Ω−1 P µm BL ¢ m ¡ −1 m0 0 −1 0 −1 0 −1 +µBL Σ + P Ω P µBL + µ Σ µ + v Ω v ¢ ¡ −µm0 Σ−1 + P0 Ω−1 P µm BL BL ¢ ¡ −1 m 0 0 −1 = (x − µBL ) Σ + P Ω P (x − µm ) + α.Using the following matrix identity (A and D invertible. BL £ −1 ¤ ¡ ¢ Σ µ + P0 Ω−1 v ≡ Σ−1 + P0 Ω−1 P µm . conformable. BL 15 (67) .4 Tedious calculations for the joint distribution of views and market Expanding the expression in square brackets in (52) we obtain: [· · · ] = (x − µ)0 Σ−1 (x − µ) + (v − Px)0 Ω−1 (v − Px) (64) = x0 Σ−1 x − 2x0 Σ−1 µ + µ0 Σ−1 µ + v0 Ω−1 v − 2x0 P0 Ω−1 v + x0 P0 Ω−1 Px ¢ ¤ ¡ £ = x0 Σ−1 + P0 Ω−1 P x − 2x0 Σ−1 µ + P0 Ω−1 v + µ0 Σ−1 µ + v0 Ω−1 v We deﬁne ¡ ¢−1 ¡ −1 ¢ µm (v) ≡ Σ−1 + P0 Ω−1 P Σ µ + P0 Ω−1 v . BL ≡ ³ ´¡ ¡ ¢−1 ¢ = Σ − ΣP0 PΣP0 + Ω PΣ Σ−1 π + P0 Ω−1 v ³ ´ ¡ ¢−1 = π + ΣP0 Ω−1 − PΣP0 + Ω PΣP0 Ω−1 v ¢−1 ¡ Pπ −ΣP0 PΣP0 + Ω (60) Noticing that ¡ ¢−1 ¡ ¢−1 Ω−1 − PΣP0 + Ω PΣP0 Ω−1 = PΣP0 + Ω . (61) ¡ ¢ which can be easily checked left-multiplying both sides by PΣP0 + Ω . using (59) we write (58) as ¡ −1 ¢−1 Σm Σ + P0 Ω−1 P BL ≡ ¢−1 ¡ = Σ − ΣP0 PΣP0 + Ω PΣ. B and C generic. potentially of diﬀerent size) ¢−1 ¡ ¢−1 ¡ = A−1 − A−1 B CA−1 B − D CA−1 (59) A − BD−1 C we can write (54) as ¡ −1 ¢−1 ¡ −1 ¢ µm Σ + P0 Ω−1 P Σ π + P0 Ω−1 v . (62) BL Similarly. we can write (60) as ¢−1 ¡ µm = π + ΣP0 PΣP0 + Ω (v − Pπ) .

= (v − v) Ω + PΣP0 ³ ¡ ¢−1 −1 ´ µ Σ φ ≡ µ0 Σ−1 − Σ−1 Σ−1 + P0 Ω−1 P ¡ ¢−1 e −e 0 Ω + PΣP0 v v 16 (75) . potentially of diﬀerent size) ¢−1 ¡ ¢−1 ¡ = A−1 − A−1 B CA−1 B − D CA−1 . we deﬁne ¡ ¢−1 −1 ¢ ¡ e Σ µ v ≡ − Ω + PΣP0 Ω−1 P Σ−1 + P0 Ω−1 P (71) (72) or equivalently ¡ ¢−1 −1 ¢−1 ¡ e Σ µ v ≡ −Ω−1 P Σ−1 + P0 Ω−1 P Ω + PΣP0 Substituting (71) and (73) back in (69) we obtain (73) where ¡ ¢−1 ¡ ¢−1 e v − 2v0 Ω + PΣP0 v (74) α = v0 Ω + PΣP0 ³ ¢−1 −1 ´ ¡ ¢−1 ¡ −1 e Σ +e 0 Ω + PΣP0 v µ v + µ0 Σ−1 − Σ−1 Σ + P0 Ω−1 P ¡ ¢−1 e −e 0 Ω + PΣP0 v v ¡ ¢−1 e e 0 (v − v) + φ. conformable. Also. A − BD−1 C (70) (69) α = µ0 Σ−1 µ + v0 Ω−1 v ¢ ¡ −1 ¢−1 ¡ −1 ¢ ¡ 0 −1 0 −1 0 −1 0 −1 Σ µ+PΩ v − µΣ +v Ω P Σ +PΩ P ¢−1 −1 ¡ −1 Σ µ = µ0 Σ−1 µ + v0 Ω−1 v − µ0 Σ−1 Σ + P0 Ω−1 P ¡ −1 ¢−1 0 −1 0 −1 0 −1 PΩ v −v Ω P Σ + P Ω P ¡ −1 ¢−1 0 −1 PΩ v +2µ0 Σ−1 Σ + P0 Ω−1 P n ¡ −1 ¢−1 0 −1 o 0 −1 −1 0 −1 = v Ω −Ω P Σ +PΩ P PΩ v ¡ ¢−1 −1 Σ µ +2v0 Ω−1 P Σ−1 + P0 Ω−1 P ³ ¢−1 −1 ´ ¡ −1 0 −1 −1 0 −1 +µ Σ − Σ Σ µ Σ +PΩ P we write the expression in curly brackets as follows: ¡ ¢−1 0 −1 ¡ ¢−1 {· · · } = Ω−1 − Ω−1 P Σ−1 + P0 Ω−1 P P Ω = Ω + PΣP0 .where Substituting the deﬁnition (65) in this expression we obtain: ¢ ¡ α ≡ µ0 Σ−1 µ + v0 Ω−1 v − µm 0 Σ−1 + P0 Ω−1 P µm BL BL (68) Using the following matrix identity (A and D invertible. B and C generic.

Substituting (74) back in (67) the expression in square brackets in (52) reads ¢ 0¡ [· · · ] = (x − µm (v)) Σ−1 + P0 Ω−1 P (x − µm (v)) (76) BL BL ¢ 0¡ 0 −1 e e + (v − v) Ω + PΣP (v − v) + φ. 2 1 |Ω| 0 −1 2 e− 2 (x−µBL (v)) (Σ 1 0 −1 m 0 −1 +P0 Ω−1 P)(x−µm (v)) BL where the last equality follows from ¯ ¯ ¯Ω + PΣP0 ¯ = 1. e− 2 (v−v) (Ω+PΣP ) (v−v) (77) ¯ −1 ¯ 1 − 1 (x−µm (v))0 (Σ−1 +P0 Ω−1 P)(x−µm (v)) BL BL = ¯Σ + P0 Ω−1 P¯ 2 e 2 −1 ¯ ¯ 1 ¯Ω + PΣP0 ¯− 2 e− 1 (v−v)0 (Ω+PΣP0 ) (v−v) .V (x. (80) which in turn follows from this identity (B and C generic. potentially of diﬀerent size): 17 . v) ∝ |Σ| −1 2 does not depend on either v or x. |Σ| |Ω| |Σ−1 + P0 Ω−1 P| (78) which in turns follows from ¯ ¯ ¡ ¯ ¢¯ |Σ| |Ω| ¯Σ−1 + P0 Ω−1 P¯ = ¯Σ Σ−1 + P0 Ω−1 P ¯ |Ω| ¯ ¯ (79) = ¯I + ΣP0 Ω−1 P¯ |Ω| ¯ ¯ ¡ ¯ ¢¯ ¯I + Ω−1 PΣP0 ¯ |Ω| = ¯Ω I + Ω−1 PΣP0 ¯ = ¯ ¯ = ¯Ω + PΣP0 ¯ . conformable. Therefore neither does φ in (75).Therefore the joint probability (52) becomes: fX. |IJ + CB| ≡ |IK + BC| .